Free Kleinian groups and volumes of hyperbolic 3-manifolds

Transcription

1 Free Kleinian groups and volumes of hyperbolic 3-manifolds James W. Anderson, Richard D. Canary, Marc Culler, and Peter B. Shalen October 1, Introduction The central result of this paper, Theorem 6.1, gives a constraint that must be satisfied by the generators of any free, topologically tame Kleinian group without parabolic elements. The following result is case (a) of Theorem 6.1. Main Theorem Let k 2 be an integer and let Φ be a purely loxodromic, topologically tame discrete subgroup of Isom + (H 3 ) which is freely generated by elements ξ 1,..., ξ k. Let z be any point of H 3 and set d i = dist(z, ξ i z) for i = 1,..., k. Then we have k i= e d i 1 2. In particular there is some i {1,..., k} such that d i log(2k 1). The last sentence of the Main Theorem, in the case k = 2, is equivalent to the main theorem of [14]. While most of the work in proving this generalization involves the extension from rank 2 to higher ranks, the main conclusion above is strictly stronger than the main theorem of [14] even in the case k = 2. partially supported by an NSF-NATO postdoctoral fellowship partially supported by a Sloan Foundation Fellowship and an NSF grant partially supported by an NSF grant partially supported by an NSF grant 1

2 1. Introduction 2 Like the main result of [14], Theorem 6.1 has applications to the study of large classes of hyperbolic 3-manifolds. This is because many subgroups of the fundamental groups of such manifolds can be shown to be free by topological arguments. The constraints on these free subgroups impose quantitative geometric constraints on the shape of a hyperbolic manifold. As in [14] these can be applied to give volume estimates for hyperbolic 3-manifolds satisfying certain topological restrictions. The volume estimates obtained here, unlike those proved in [14], are strong enough to have qualitative consequences, as we shall explain below. The following result is proved by combining the case k = 3 of the Main Theorem with the techniques of [15]. Corollary 9.2 Let N be a closed orientable hyperbolic 3-manifold. Suppose that the first betti number β 1 (N) is at least 4, and that π 1 (N) has no subgroup isomorphic to the fundamental group of a surface of genus 2. Then N contains a hyperbolic ball of radius 1 log 5. Hence the volume of N is greater than There is no reason to expect these estimates to be sharp. For instance, empirical evidence based on Weeks census [39] suggests that the conclusion of the corollary may hold under the hypothesis that β 1 (N) is at least 2, with no assumption on the surface subgroups of π 1 (N). However, the significance of our results lies elsewhere. The point is that these results imply that certain topological conditions on the manifold follow from an upper bound on the volume. More specifically, the volumes of hyperbolic 3-manifolds are known to form a well-ordered set of ordinal type ω ω. If one lists the closed hyperbolic manifolds in ascending order of volume, the topological complexity of the manifolds tends to grow as one progresses through the list. We are interested in understanding this phenomenon in an explicit way. The above result provides explicit information of this type. The volume of a cusped manifold is larger than that of any of its Dehn fillings, and is a limit point of the set of volumes of such fillings. There are 8 distinct volumes less than 3.08 among the volumes of orientable cusped manifolds in the Weeks census. Thus the result implies that each of the manifolds realizing the first 8ω volumes either has betti number at most 3 or has a fundamental group containing an isomorphic copy of a genus-2 surface group. (This conclusion is stated as Corollary 9.3.) It was not possible to deduce qualitative consequences of this sort in [14] because the lower bound of 0.92, obtained there for the

3 1. Introduction 3 volume of a closed hyperbolic 3-manifold of first betti number at least 3, is smaller than the least known volume of any hyperbolic 3-manifold. Corollary 9.4 is similar to the above corollary but illustrates the applicability of our techniques to the geometric study of infinite-volume hyperbolic 3-manifolds. It asserts that a non-compact, topologically tame, orientable hyperbolic 3-manifold N without cusps always contains a hyperbolic ball of radius 1 log 5 unless π 2 1(N) either is a free group of rank 2 or contains an isomorphic copy of a genus-2 surface group. Another application of Theorem 6.1 to non-compact finite-volume manifolds is the following result, which uses only the case k = 2 of the Main Theorem, but does not follow from the weaker form of the conclusion which appeared in [14]. Theorem 11.1 Let N = H 3 /Γ be a non-compact hyperbolic 3-manifold. If N has betti number at least 4, then N has volume at least π. Theorem 11.1 is deduced via Dehn surgery techniques from Proposition 10.1 and its Corollary 10.3, which are of independent interest. These results imply that if a hyperbolic 3-manifold satisfies certain topological restrictions, for example if its first betti number is at least 3, then there is a good lower bound for the radius of a tube about a short geodesic, from which one can deduce a lower bound for the volume of the manifold in terms of the length of a short geodesic. This lower bound approaches π as the length of the shortest geodesic tends to 0. Corollary 10.3 will be used in [13] as one ingredient in a proof of a new lower bound for the volume of a hyperbolic 3-manifold of betti number 3. This lower bound is greater than the smallest known volume of a hyperbolic 3-manifold, and therefore has the qualitative consequence that any smallest-volume hyperbolic 3-manifold has betti number at most 2. The proof of the Main Theorem follows the same basic strategy as the proof of the main theorem of [14]. The Main Theorem is deduced from Theorem 6.1(d), which gives the same conclusion under somewhat different hypotheses. In 6.1(d), rather than assuming that the free Kleinian group Φ is topologically tame and has no parabolics, we assume that the manifold H 3 /Φ admits no non-constant positive superharmonic functions. As in [14], the estimate is proved in this case by using a Banach-Tarski-style decomposition of the area measure based on a Patterson construction. The deduction of the Main Theorem from 6.1(d) is based on Theorem 5.2, which asserts that, in the variety of representations of a free group F k, the boundary of the set CC(F k ) of convex-

4 1. Introduction 4 cocompact discrete faithful representations contains a dense G δ consisting of representations whose images are analytically tame Kleinian groups without parabolics. This was proved in [14] in the case k = 2. By definition, a rank-k free Kleinian group Γ without parabolics is analytically tame if the convex core of H 3 /Γ can be exhausted by a sequence of geometrically well behaved compact submanifolds (a more exact definition is given in Section 5). The case k = 2 of Theorem 5.2 was established in [14] by combining a theorem of McMullen s [29] on the density of maximal cusps on the boundary of CC(F k ) with a special argument involving the canonical involution of a 2-generator Kleinian group. The arguments used in the proof of Theorem 5.2 make no use of the involution. This makes possible the generalization to arbitrary k, while also giving a new proof in the case k = 2. The ideas needed for the proof are developed in sections 2 through 5, and will be sketched here. In Section 3 we prove a general fact, Proposition 3.2, about a sequence (ρ n ) of discrete faithful representations of a finitely generated, torsion-free, non-abelian group G which converges to a maximal cusp ω. (For our purposes a maximal cusp is a discrete faithful representation ω of G into PSL 2 (C) such that ω(g) is geometrically finite and every boundary component of the convex core of H 3 /ω(g) is a thrice-punctured sphere.) After passing to a subsequence one can assume that the Kleinian groups ρ n (G) converge geometrically to a Kleinian group Γ, which necessarily contains ω(g) as a subgroup. Proposition 3.2 then asserts that the convex core of N = H 3 /ω(g) embeds isometrically in H 3 / Γ. To prove this, we use Proposition 2.7, which combines an algebraic characterization of how conjugates of ω(g) can intersect in the geometric limit (Lemma 2.4), and a description of the intersection of the limit sets of two topologically tame subgroups of a Kleinian group (Theorem 2.5). In Section 4 we construct a large submanifold D of the convex core of N which is geometrically well-behaved in the sense that D has bounded area and the radius-2 neighborhood of D has bounded volume. We use Proposition 3.2 to show that if ρ is a discrete faithful representation near enough to ω, then H 3 /ρ(g) contains a nearly isometric copy of D. This copy is itself geometrically well-behaved in the same sense. In Section 5 we specialize to the case G = F k. We show that if a discrete faithful representation ρ is well-approximated by infinitely many maximal cusps, then its associated quotient manifold contains infinitely many geometrically well-behaved submanifolds. In fact, we show that the resulting

5 1. Introduction 5 submanifolds exhaust the convex core of the quotient manifold and hence that the quotient manifold is analytically tame. We then apply McMullen s theorem to prove that there is a dense G δ in the boundary of CC(F k ) consisting of representations which can be well approximated by maximal cusps. In the argument given in [14], the involution of a 2-generator Kleinian group is used not only in the deformation argument, but also in the calculation based on the decomposition of the area measure in the case where H 3 /Φ supports no non-constant superharmonic functions. The absence of an involution in the k- generator case is compensated for by a new argument based on the elementary inequality established in Lemma 6.2. This leads to the stronger conclusion of the main theorem in the case k = 2. Section 6 is devoted to the proof of Theorem 6.1. We have mentioned that the application of Theorem 6.1 to the geometry of hyperbolic manifolds depends on a criterion for subgroups of fundamental groups of such manifolds to be free. The first such criterion in the case of a 2-generator subgroup was proved in [19] and independently in [37]. A partial generalization to k-generator subgroups, applying only when the given manifold is closed, was given in [3]. In Section 7 of this paper we give a criterion that includes the above results as special cases and is adapted to the applications in this paper. In Section 8 we introduce a generalization of the notion of a Margulis number. We say that a positive number λ is a k-margulis number for a Kleinian group Γ if the following condition holds: if ξ 1,..., ξ k are elements of Γ and if there exists a point z H 3 which is displaced less than λ by each ξ i then the group ξ 1,..., ξ k can be generated by k 1 abelian subgroups. In the case k = 2 the group ξ 1, ξ 2 would have to be abelian; thus the notion of a 2-Margulis number coincides with that of a Margulis number as defined in [14] and [33]. This notion, and the related notion of a k-strong Margulis number proves useful for organizing the applications of the results of the earlier sections to the study of hyperbolic manifolds. The applications are presented in Sections 9, 10 and 11. The pictures of limit sets of maximal cusps which appear in Section 3 were created by Yair Minsky, and were based on some earlier pictures by Chris Bishop. We are grateful to them for allowing us to use them here. We close the introduction by mentioning a few notational conventions which are used throughout. We use H G to denote that H is a subgroup of G, and H < G to denote that H is a proper subgroup of G. The translate of

6 2. On the sphere at infinity 6 a set X by a group element γ is denoted γ X. Finally, we use dist(z, w) to denote the hyperbolic distance between points z and w in H 3. 2 On the sphere at infinity In this section we introduce the notion of the geometric limit of a sequence of Kleinian groups. We will consider a convergent sequence of discrete faithful representations into PSL 2 (C) whose images converge geometrically. In general, the geometric limit of the images contains the image of the limit as a subgroup. The results in this section characterize the intersection of two conjugates of this subgroup, and the intersection of their limit sets. The group PSL 2 (C) will be considered to act either by isometries on H 3 or, via extension to the sphere at infinity, by Möbius transformations on the Riemann sphere C. The action of a discrete subgroup Γ of PSL 2 (C) partitions C into two pieces, the domain of discontinuity Ω(Γ) and the limit set Λ(Γ). The domain of discontinuity is the largest open subset of C on which Γ acts properly discontinuously. If Λ(Γ) contains two or fewer points, we say Γ is elementary. If Γ has an invariant circle in C and preserves an orientation of the circle then we say that Γ is Fuchsian. By a Kleinian group we will mean a discrete non-elementary subgroup Γ of PSL 2 (C). We will say that a Kleinian group Γ is purely parabolic if every nontrivial element is parabolic, or purely loxodromic if every non-trivial element is loxodromic. Given a finitely generated group G, let Hom(G, PSL 2 (C)) denote the variety of representations of G into PSL 2 (C). A choice of k elements which generate G determines a bijection from Hom(G, PSL 2 (C)) to an algebraic subset of (PSL 2 (C)) k. We give Hom(G, PSL 2 (C)) the topology that makes this bijection a homeomorphism onto the algebraic set with its complex topology. This topology on Hom(G, PSL 2 (C)) is independent of the choice of generators of G. For the rest of this section, and throughout section 3, we will assume that G is a finitely generated, non-abelian, torsion-free group. Let D(G) denote the subspace of Hom(G, PSL 2 (C)) which consists of those representations which are injective and have discrete image. It is a fundamental result of Jørgensen s [20] that D(G) is a closed subset of Hom(G, PSL 2 (C)). The proof of Jørgensen s result is based on an inequality for discrete subgroups

7 2. On the sphere at infinity 7 of PSL 2 (C). A second consequence of this inequality is the following lemma. Lemma 2.1 (Lemma 3.6 in [21]) Let (ρ n ) be a convergent sequence in D(G). If (g n ) is a sequence of elements of G such that (ρ n (g n )) converges to the identity, then there exists n 0 such that g n = 1 for n n A sequence of discrete subgroups (Γ n ) is said to converge geometrically to a Kleinian group Γ if and only if (1) for every γ Γ, there exist elements γ n Γ n such that the sequence (γ n ) converges to γ, and (2) whenever (Γ nj ) is a subsequence of (Γ n ) and γ nj Γ nj are elements such that the sequence (γ nj ) converges to a Möbius transformation γ, we have γ Γ. We call Γ the geometric limit of (Γ n ). The following basic fact is proved in Jørgensen-Marden [21]. Proposition 2.2 (Proposition 3.8 in [21]) Let (ρ n ) be a sequence of elements of D(G) converging to ρ. Then (ρ n (G)) has a geometrically convergent subsequence. If Γ is the geometric limit of any such subsequence, then ρ(g) Γ. The following fact will also be used. Lemma 2.3 Let (ρ n ) be a convergent sequence in D(G) such that (ρ n (G)) converges geometrically to Γ. Then Γ is torsion-free. Proof of 2.3: Suppose that γ Γ has finite order d. Let (g n ) be a sequence of elements of G such that ρ n (g n ) converges to γ. Then ρ n (gn) d converges to the identity. Hence by Lemma 2.1 we have gn d = 1 for large n. Since G is torsion-free, we have g n = 1 for large n and therefore γ =

8 2. On the sphere at infinity 8 The following lemma characterizes the intersection of two conjugates of ρ(g) in the geometric limit Γ. Lemma 2.4 Let (ρ n ) be a sequence of elements of D(G) converging to ρ. Suppose that the groups ρ n (G) converge geometrically to Γ. Then ρ(g) γρ(g)γ 1 is a (possibly trivial) purely parabolic group for each γ Γ ρ(g). Proof of 2.4: Suppose that ρ(a) is a nontrivial element of ρ(g) γρ(g)γ 1 for some γ Γ ρ(g). We will show that ρ(a) is parabolic. Assume to the contrary that ρ(a) is loxodromic. We may write ρ(a) = γρ(b)γ 1 for some b G. Choose g n G such that (ρ n (g n )) converges to γ. We then have that (ρ n (g n bgn 1 )) converges to ρ(a), and hence that (ρ n (a 1 g n bgn 1 )) converges to 1. It follows from Lemma 2.1 that there exists an integer n 0 such that a = g n bgn 1 for all n n 0. Hence gn 1 0 g n is contained in the centralizer of b for all n n 0. Applying ρ n and passing to the limit, we have that ρ(gn 1 0 )γ commutes with ρ(b). Since the Kleinian group Γ is torsion-free by Lemma 2.3 and the element ρ(b) Γ is loxodromic, the centralizer of ρ(b) in Γ is cyclic. Thus there are integers j and k such that (ρ(gn 1 0 )γ) j = ρ(b) k. A second application of Lemma 2.1 shows that for some n 1 n 0 we have (gn 1 0 g n ) j = b k for all n n 1. But since G is isomorphic to a torsion-free Kleinian group, each element of G has at most one j th root. Hence b k has a unique j th root c and we have gn 1 0 g n = c for n M. Thus g n = g n0 c for large n, so the sequence (g n ) is eventually constant. Since (ρ n (g n )) converges to γ, this implies that γ is an element of ρ(g), which contradicts our hypothesis that γ Γ ρ(g). 2.4 Next we consider the intersection of the limit set of ρ(g) with its image under an element of the geometric limit Γ. The following definition will be useful. Let Γ 1 and Γ 2 be subgroups of the Kleinian group Γ. We will say that a point p Λ(Γ 1 ) Λ(Γ 2 ) is in P(Γ 1, Γ 2 ) if and only if Stab Γ1 (p) = Z, Stab Γ2 (p) = Z, and Stab Γ1 (p), Stab Γ2 (p) = Z Z. In particular, it must be that p is a parabolic fixed point of both Γ 1 and Γ 2. We will make use of the following result, due to Soma [34] and Anderson [1], which provides the link between the intersection of the limit sets of a pair of subgroups and the limit set of the intersection of the subgroups.

9 2. On the sphere at infinity 9 Recall that a Kleinian group Γ is topologically tame if H 3 /Γ is homeomorphic to the interior of a compact 3-manifold. Theorem 2.5 Let Γ 1 and Γ 2 be nonelementary, topologically tame subgroups of the Kleinian group Γ. Then Λ(Γ 1 ) Λ(Γ 2 ) = Λ(Γ 1 Γ 2 ) P(Γ 1, Γ 2 ). The next lemma shows that the term P(Γ 1, Γ 2 ) may be ignored in the case where Γ 1 and Γ 2 are distinct conjugates of ρ(g) by elements of the geometric limit Γ. Lemma 2.6 Let (ρ n ) be a sequence in D(G) converging to ρ. Suppose that the groups ρ n (G) converge geometrically to Γ. Then for each γ Γ ρ(g), the set P(ρ(G), γρ(g)γ 1 ) is empty. Proof of 2.6: The argument runs along much the same line as the proof of Proposition 2.4. Suppose that p P(ρ(G), γρ(g)γ 1 ), that Stab ρ(g) (p) = Z is generated by ρ(a) and that Stab γρ(g)γ 1(p) = Z is generated by γρ(b)γ 1. Note that each of the elements a and b generates its own centralizer. Choose g n G so that (ρ n (g n )) converges to γ. Since ρ(a) commutes with γρ(b)γ 1, we conclude as in the proof of Lemma 2.4 that there exists an integer n 0 such that a commutes with g n bgn 1 for all n n 0. Since a generates its centralizer in G, each of the elements g n bgn 1 for n n 0 must be a power of a. But these are all conjugate elements, while distinct powers of a are not conjugate. Therefore we must have that g n bgn 1 = g n 0 bgn 1 0 for all n n 0. Thus gn 1 0 g n commutes with b for all n n 0. Since the centralizer of b is cyclic, we may now argue exactly as in the proof of 2.4 that the sequence (g n ) must be constant for n n 1 n 0, obtaining a contradiction to our hypothesis that γ Γ ρ(g). 2.6 As an immediate consequence of Lemma 2.4, Theorem 2.5, and Lemma 2.6, we have the following proposition. Proposition 2.7 Let (ρ n ) be a sequence in D(G) converging to ρ. Suppose that the groups ρ n (G) converge geometrically to Γ and that ρ(g) is topologically

10 3. In the convex core 10 tame. Then for any γ Γ ρ(g) the group ρ(g) γρ(g)γ 1 is purely parabolic and Λ(ρ(G)) γ Λ(ρ(G)) = Λ(ρ(G) γρ(g)γ 1 ). Hence if Λ(ρ(G)) γ Λ(ρ(G)) is non-empty then it must contain only the fixed point of ρ(g) γρ(g)γ 1. 3 In the convex core We continue to assume that G is a finitely generated nonabelian torsion-free group. We consider a sequence (ρ n ) in D(G) which converges to a representation ω which is a maximal cusp (defined below). If we assume that the groups ρ n (G) converge geometrically to Γ then the hyperbolic manifold N = H 3 /ω(g) is a covering space of N = H 3 / Γ. The main result of this section says that in this situation the restriction of the covering projection gives an embedding of the convex core of N into N. It may be helpful to recall the most basic situation in which the convex core of a manifold does not embed in a manifold which it covers. Let N be a hyperbolic 3-manifold and let f : S N be a totally geodesic isometric immersion of a finite area surface S. Let N be the cover of N associated to π 1 (S), so that f lifts to a totally geodesic embedding f : S N. Since f(s) is totally geodesic, Λ(π 1 (S)) is a circle and f(s) is the convex core of N. The convex core of N embeds in N if and only if f is an embedding. Notice that f is not an embedding if and only if there exists an element γ of π 1 ( N) such that γ(λ(π 1 (S))) intersects Λ(π 1 (S)) transversely, hence in at least two points. We will be dealing with the case where the algebraic limit is a maximal cusp and hence each boundary component of the convex core of our algebraic limit is a totally geodesic thrice-punctured sphere (Lemma 3.1). We will see, as in the example above, that if the convex core of the algebraic limit does not embed then there must be an element γ of Γ ω(g) such that γ(λ(ω(g))) intersects Λ(ω(G)) in at least two points. An application of Proposition 2.7 will complete the proof. Given a Kleinian group Γ, define its convex hull CH(Γ) in H 3 to be the smallest non-empty convex set in H 3 which is invariant under the action of Γ. Thus CH(Γ) is the intersection of all half-spaces in H 3 whose closures in the compactification H 3 C contain Λ(Γ). (Recall that a Kleinian group is, by definition, non-elementary so that its limit set has more than two points.)

11 3. In the convex core 11 The convex core of N = H 3 /Γ is C(N) = CH(Γ)/Γ. We say that N, or equivalently Γ, is geometrically finite if Γ is finitely generated and C(N) has finite volume. The injectivity radius inj N (x) of N at the point x is half the length of the shortest homotopically non-trivial closed loop passing through x. Note that injectivity radius increases under lifting to a covering space. That is, if N is a cover of N with covering map π : N N, and x is any point of N, then inj N (x) inj N(π(x)). Given a hyperbolic 3-manifold N, define the ɛ-thick part of N as and the ɛ-thin part of N as N thick(ɛ) = {x N inj N (x) ɛ 2 } N thin(ɛ) = {x N inj N (x) ɛ 2 }. We recall that C(N) N thick(ɛ) is compact for all ɛ > 0 if and only if N is geometrically finite (see Bowditch [7]). Hence, for a geometrically finite hyperbolic 3-manifold N, the sets C(N) N thick(1/m) for m 1 form an exhaustion of C(N) by compact subsets. We say that a representation ω in D(G) is a maximal cusp if N = H 3 /ω(g) is geometrically finite and every component of the boundary C(N) of its convex core is a thrice-punctured sphere. We will further require that ω(g) not be a Fuchsian group. (This rules out only the case that ω(g) is in the (unique) conjugacy class of finite co-area Fuchsian groups uniformizing the thrice-punctured sphere.) Maximal cusps are discussed at length by Keen, Maskit and Series in [22], where the image groups are termed maximally parabolic. A proof of the following lemma appears in [22]; since we will be using the lemma heavily, we include a sketch of the proof here. Lemma 3.1 Let ω D(G) be a maximal cusp, and let N = H 3 /ω(g). Then each component of C(N) is totally geodesic. Proof of 3.1: Since the universal cover of C(N) is CH(ω(G)), it suffices to show that each component of CH(ω(G)) is a totally geodesic hyperplane or, equivalently, that each component of Ω(ω(G)) is a disk bounded by a circle on the sphere at infinity.

12 3. In the convex core 12 Figure 1: The domain of discontinuity of a maximal cusp is a union of round disks Recall, for example from Epstein-Marden [17], that C(N) is homeomorphic to Ω(ω(G))/ω(G). Moreover, by the Ahlfor s Finiteness theorem Ω(ω(G))/ω(G) has finite area. Thus each component S of Ω(ω(G))/ω(G) must be a thrice-punctured sphere. Write S = /Γ, where is a component of Ω(ω(G)) and Γ is the subgroup of ω(g) stabilizing. Since /Γ is a thrice-punctured sphere, the group Γ must be a Fuchsian group and must be a disk bounded by a circle on the sphere at infinity. (For a proof, see Chapter IX.C of Maskit s book [26]; for a picture see figure 1 below.) 3.1 We are now ready to prove the main result of this section. A map between locally compact spaces will be called an embedding if it is proper and one-toone. Proposition 3.2 Let (ρ n ) be a sequence of elements of D(G) converging to a maximal cusp ω. Suppose that the groups ρ n (G) converge geometrically to Γ.

13 3. In the convex core 13 Let N = H 3 /ω(g), N = H 3 / Γ, and let π : N N be the covering map. Then π C(N) is an embedding. Figure 2: The limit set of a maximal cusp may intersect its translate in at most one point. Proof of 3.2: We first note that if (x i ) is a sequence in C(N) leaving every compact set, then lim i inj N (x i ) = 0. Hence, lim i inj N(π(x i )) = 0, which implies that (π(x i )) leaves every compact subset of N. Thus, π C(N) is a proper mapping. It remains to show that π is injective. The universal cover of C(N) is CH(ω(G)). Thus it suffices to show that CH(ω(G)) γ CH(ω(G)) is empty for each γ Γ ω(g). For notational convenience, set X = CH(ω(G)). Since ω is assumed to be a maximal cusp, each component of C(N) is totally geodesic. Hence each component of X is a plane H in H 3 whose boundary at infinity is a circle C which lies in Λ(ω(G)). If X γ X is not empty, there are two possibilities. Either there is a point in X γ X, or a component of X lies entirely within γ X (or vice versa). We begin with the case that there is a point x in X γ X. There then exist a plane H in X and a plane H in γ X with x H H. If two planes

14 3. In the convex core 14 Figure 3: The convex hull of the limit set of a maximal cusp cannot intersect its translate in H 3 meet, either their intersection is a line or they are equal. If we let C denote the boundary at infinity of H and C the boundary at infinity of H, then either C C contains exactly two points or C = C. However, C C is contained in Λ(ω(G)) γ Λ(ω(G)), so that Λ(ω(G)) γ Λ(ω(G)) contains at least two points, which contradicts Proposition 2.7. The second possibility is that a component H of X lies entirely within γ X. In this case, the boundary at infinity C of H lies in the boundary at infinity of γ X, which is exactly γ Λ(ω(G)). However, C also lies in Λ(ω(G)); hence Λ(ω(G)) γ (Λ(ω(G))) contains C. This also contradicts Proposition 2.7. See figures 2 and 3 for an illustration of how Λ(ω(G)) can meet its translate by γ. 3.2 Remark 3.3 The conclusion of Proposition 3.2 holds, by the same argument, whenever N is topologically tame and C(N) is totally geodesic. In general, the convex core of the algebraic limit need not embed in the geometric limit. However, one can define the visual hull of a Kleinian group Γ to be the set of all points in H 3 such that the visual area of every component of Ω(Γ) is at most 1. The visual core is then defined to be the quotient of the 2 visual hull. Notice that if Γ is a maximal cusp, then its visual core and its convex core coincide.

15 4. Near a maximal cusp 15 Anderson and Canary [2] use techniques similar to those developed in this section to prove that the visual core embeds whenever the algebraic limit has connected limit set and no accidental parabolics. They also show that, under the same assumptions, there is a compact core for the algebraic limit which embeds in the geometric limit. It has recently been discovered that the visual core of the algebraic limit need not embed in the algebraic limit even if the algebraic limit has connected limit set. 4 Near a maximal cusp In this section, we will prove that if a representation in D(G) is near enough to a maximal cusp, then its associated hyperbolic 3-manifold contains a nearly isometric copy of an ɛ-truncated convex core of the maximal cusp. We first define this ɛ-truncated object and describe some of its useful attributes. We recall that it follows from the Margulis lemma that there exists a constant λ 0, such that if ɛ < λ 0 and N is a hyperbolic 3-manifold, then every component P of N thin(ɛ) is either a solid torus neighborhood of a closed geodesic, or the quotient of a horoball H by a group Θ of parabolic elements fixing H (see for example [4]). In the second case, Θ is isomorphic either to Z or to Z Z. Moreover, H is precisely invariant under Θ < Γ, by which we mean that if γ Γ and γ H H, then γ Θ and γ H = H. If Θ = Z, we call P a rank-one cusp, and if Θ = Z Z, we call P a rank-two cusp. Recall also that there exists L(ɛ) > 0, such that any two components of N thin(ɛ) are separated by a distance of at least L(ɛ). The next lemma gives the structure of N thin(ɛ) for sufficiently small ɛ. For a proof see section 6 of Morgan [31]. Lemma 4.1 Let N be a geometrically finite hyperbolic 3-manifold. There exists δ(n) < λ 0, such that if ɛ δ(n) and P is a component of N thin(ɛ), then (i) P is non-compact, (ii) P meets C(N) orthogonally along each component of their intersection. (iii) E = P C(N) is a Euclidean surface with geodesic boundary, and diam E 1; (iv) if P is a rank-one cusp then E is an annulus, and if P is a rank-two cusp then E is a torus; and

16 4. Near a maximal cusp 16 (v) C(N) P is homeomorphic to E [0, ). In particular, for any ɛ δ(n) the set N thick(ɛ) C(N) is a compact 3-manifold with piecewise smooth boundary. Given a geometrically finite hyperbolic 3-manifold N we may define its ɛ-truncated convex core D ɛ (N) to be the intersection of its convex core C(N) with the ɛ-thick part N thick(ɛ) of N. The above lemma completely characterizes C(N) D ɛ (N) when ɛ < δ(n). Recall that a compact submanifold of N is said to be a compact core for N if the inclusion map is a homotopy equivalence. Lemma 4.2 Suppose that N is a geometrically finite hyperbolic 3-manifold, and δ(n) > ɛ > 0. Then D ɛ (N) is a compact core for N. Proof of 4.2: First recall that the inclusion of C(N) into N is a homotopy equivalence. Moreover, each component of C(N) D ɛ (N) is homeomorphic to E (0, ) for some Euclidean surface E. Thus, the inclusion of D ɛ (N) into C(N) is a homotopy equivalence. The result follows. 4.2 For any subset X of a hyperbolic manifold N we will denote by N r (X) the closed neighborhood of radius r of X. In the case N = H 3 /ω(g), where ω is a maximal cusp, Proposition 4.4 will provide bounds for both the area of D ɛ (N) and the volume of N 2 ( D ɛ (N)). These bounds will depend only on the topological type of N and not on ɛ. We first recall the following special case of Lemma 8.2 in [8] (see also Proposition of Thurston [36]). Lemma 4.3 There is a constant κ > 0, such that for any maximal cusp ω D(G) and any collection S of components of the boundary of the convex core of N = H 3 /ω(g), the neighborhood N 3 (S) has volume less than κ area S. 4.3 If N = H 3 /ω(g), where ω(g) is a maximal cusp, we will denote by σ(n) the number of components of C(N), and by τ(n) the number of rank-two cusps of N. We set α(n) = 7 πσ(n) + 2πτ(N) 2

17 4. Near a maximal cusp 17 and β(n) = 2πκσ(N) + πe 4 τ(n), where κ is the constant given by Lemma 4.3. Lemma 4.4 Let ω be a maximal cusp and let N = H 3 /ω(g). If δ(n) > ɛ > 0, then (i) area D ɛ (N) α(n), and (ii) vol(n 2 ( D ɛ (N))) β(n). Proof of 4.4: Notice that D ɛ (N) = S E where S = C(N) N thick(ɛ) and E = C(N) N thick(ɛ). Since S C(N) and since each component of C(N) is a thrice-punctured sphere, we have area S area C(N) = 2πσ(N). By Lemma 4.1, each component of E is a Euclidean manifold of diameter at most 1, so each component of E has area at most π. Since each component of C(N) contains three components of E there are 3 σ(n) annular components 2 of E. Moreover, there are τ(n) toroidal components of E. The first assertion follows. Let Ŝ denote the union of S with the annular components of E. Since each annular component of E has diameter less than 1, Thus Lemma 4.3 guarantees that N 2 (Ŝ) N 3(S) N 3 ( C(N)). vol N 2 (Ŝ) κ area( C(N)) = 2πκσ(N). Now, if T is a toroidal component of E, then N 2 (T ) is (the quotient of) a region isometric to T ( 2, 2) with the metric ds 2 = e 2t ds 2 T + dt2, which has volume less than 2πe 4. The second assertion now follows. 4.4 Our next result, Proposition 4.5, asserts (among other things) that the hyperbolic manifold associated to a representation which is near enough to a maximal cusp contains a bilipschitz copy of the ɛ-truncated convex core of the manifold associated to the maximal cusp.

18 4. Near a maximal cusp 18 We first outline the argument. Suppose that (ρ n ) is a sequence of representations in D(G) which converges to a maximal cusp ω and that the groups ρ n (G) converge geometrically to Γ. Let N = H 3 /ω(g) and N = H 3 / Γ. If π : N N is the covering map, Proposition 3.2 implies that π Dɛ(N) is an embedding. Since (ρ n (G)) converges geometrically to Γ, larger and larger chunks of N are nearly isometric to larger and larger chunks of N n = H 3 /ρ n (G). In particular, for all large enough n there exists a 2-biLipschitz embedding f n : V n N n where π(d ɛ (N)) V n N. The desired bilipschitz copy of D ɛ (N) is f n (π(d ɛ (N))). In order to carry out the program outlined above, it will be necessary to make consistent choices of base points in different hyperbolic 3-manifolds. We will use the following convention. If z is a point in H 3 and Γ is a Kleinian group, we will let z Γ denote the image of z in the hyperbolic manifold H 3 /Γ. In the case that Γ = ρ(g) for some representation ρ Hom(G, PSL 2 (C)) we will write z ρ = z ρ(g). If a codimension-0 submanifold X of a hyperbolic manifold N is connected and has piecewise smooth boundary, then it has two natural distance functions. In the extrinsic metric the distance between two points of X is equal to their distance in N, while in the intrinsic metric the distance is the infimum of the lengths of rectifiable paths in X joining the two points. Observe that if X and Y are submanifolds of hyperbolic manifolds and if f: X Y is a K-biLipschitz map with respect to the extrinsic metrics, then f is also K-biLipschitz with respect to the intrinsic metrics. Proposition 4.5 Suppose that ω D(G) is a maximal cusp and set N = H 3 /ω(g). Let ɛ > 0 be given with the property that ɛ < δ(n) and let z be a point of H 3 such that z ω lies in the interior of D ɛ (N). Then there is a neighborhood U(ɛ, z, ω) of ω in D(G) such that for each ρ U(ɛ, z, ω), there exists a map φ : D ɛ (N) N = H 3 /ρ(g), with the following properties: (1) φ maps D ɛ (N) homeomorphically onto a manifold with piecewise smooth boundary, and is 2-biLipschitz with respect to the intrinsic metrics on D ɛ (N) and φ(d ɛ (N)), (2) φ(z ω ) = z ρ, (3) vol N 1 ( (φ(d ɛ (N)))) 8β(N), and

19 4. Near a maximal cusp 19 (4) φ(n thin(δ) D ɛ (N)) N thin(2δ) for any δ < λ 0 2 (where λ 0 is the Margulis constant). Proof of 4.5: Let (ρ n ) be a sequence in D(G) that converges to ω, and set N n = H 3 /ρ n (G). It suffices to prove that (ρ n ) has a subsequence (ρ ni ) such that there exist maps φ i : D ɛ (N) N ni which have properties (1) (4). Given any sequence (ρ n ) in D(G) converging to ω, Proposition 2.2 guarantees that there exists a subsequence (ρ ni (G)) of (ρ n (G)) which converges geometrically to a group Γ such that ω(g) Γ. Let N = H 3 / Γ and let π : N N be the associated covering map. Proposition 3.2 guarantees that π C(N) is an embedding and hence that π Dɛ(N) is an embedding. Let D = π(d ɛ (N)). Since π is a local isometry we find using Lemma 4.4 that vol N 2 ( D) vol N 2 ( D ɛ (N)) β(n). Since (ρ ni (G)) converges geometrically to Γ, it follows from Corollary in [10] or Theorem E.1.13 in [4] that there exist smooth submanifolds V i N, numbers r i and α i, and maps f i : V i N ni such that (i) V i contains B(r i, z Γ), the closed radius-r i neighborhood of z Γ, (ii) f i (z Γ) = z ρni, (iii) r i converges to, and α i converges to 1, (iv) f i maps V i diffeomorphically onto f(v i ) and is α i -bilipschitz with respect to the extrinsic metrics on V i and f(v i ). Choose d so that D B(d, z Γ). Set µ = max{1, λ 0 /2}. We may assume that the subsequence (ρ ni ) has been chosen so that α i < 2 and r i > d + 2µ for all i. This condition on r i implies that N 2µ (D) is contained in the interior of V i. We claim that N µ (f i (D)) f i (V i ). To prove the claim, we consider the frontier X of N 2µ (D) in N and the frontier Y i of f i (N 2µ (D)) in N ni. Since f i is a homeomorphism onto its image it follows from invariance of domain is extrinsically α i -Lipschitz with α i < 2, and since every point of X has distance 2µ from D, every point of f i (X) = Y i must be a distance greater than µ from f i (D). Thus Y i is disjoint from N µ (f i (D)). Since that f i (X) = Y i. Since f 1 i

20 4. Near a maximal cusp 20 f i (N 2µ (D)) contains f i (D) and is disjoint from the frontier Y i of N µ (f i (D)) we have f i (N 2µ (D)) N µ (f i (D)), and the claim follows. In particular, N 1 (f i ( D)) f i (V i ). Again using that fi 1 is extrinsically 2-Lipschitz we conclude that N 1 (f i ( D)) f i (N 2 ( D)). On the other hand, since f i is extrinsically 2-Lipschitz, it is intrinsically 2-Lipschitz and can therefore increase volume by at most a factor of 8. Therefore vol N 1 ( f i (D)) 8 vol N 2 ( D) 8β(N). (1) We now define φ i : D ɛ (N) N ni to be f i π. We will complete the proof of the proposition by showing that φ i has properties (1) (4). Since π is a local isometry and π Dɛ(N) is an embedding, the map π Dɛ(N) is an isometry between D ɛ (N) and π(d ɛ (N)) with respect to their intrinsic metrics. Since f i : D f i (D) is an extrinsically (and hence intrinsically) 2- bilipschitz homeomorphism, it follows that φ i has property (1). We have φ i (z ω ) = f i (π(z ω )) = f i (z Γ) = z ρni. This is property (2). Property (3) follows from equation (1) since φ i (D ɛ (N)) = f i (D). It remains to check property (4). Suppose that x N thin(δ) D ɛ (N) and δ < λ 0 2. Then there exists a homotopically non-trivial loop C (in N) based at x and having length at most δ. Notice that φ i (C) has length at most 2δ. Hence φ i (x) must lie in the 2δ-thin part of N ni unless φ i (C) is homotopically trivial. But since φ i (C) has length at most 2δ, it is contained in the closed δ-neighborhood of φ i (x) in N ni. Thus if φ i (C) were homotopically trivial in N, it would lift to a loop in a ball of radius δ in H 3 whose center projects to φ i (x). Hence φ i (C) would be null-homotopic in B(δ, φ i (x)) N δ (φ i (D)) N µ (φ i (D)) f i (V i ). This would imply that fi 1 (φ i (C)) = π(c) is homotopically trivial in N, in contradiction to the fact that π is a covering map. Thus, φ i has property (4), and the proof is complete. 4.5 Remark 4.6 Lemmas 4.3 and 4.4 have analogues for general geometrically finite hyperbolic 3-manifolds, but the constants would also depend on the minimal length of a compressible curve in C(N). Proposition 4.5 remains true, by similar arguments, whenever N is geometrically finite and every component of C(N) is totally geodesic. One may use arguments similar to those in section 3 of [11] to prove that, if ρ is sufficiently near to a maximal cusp ω, then φ(d ɛ (N ω )) is a compact core for N ρ.

21 5. All over the boundary of Schottky space 21 5 All over the boundary of Schottky space We now restrict to the case where G is the free group F k on k generators, where k 2. We set D k = D(F k ). Recall that D k is a closed subset of Hom(F k, PSL 2 (C)). Let CC k denote the subset of D k consisting of representations which are convex-cocompact, i.e. are geometrically finite and have purely loxodromic image. Moreover, let B k = CC k CC k D k. It is known (see Marden [24]) that CC k is an open subset of Hom(F k, PSL 2 (C)). (The quotient of CC k under the action of PSL 2 (C) is often called Schottky space.) Let M k denote the set of maximal cusps in D k. It is a theorem of Maskit s [28] that M k B k. McMullen has further proved that M k is a dense subset of B k. This result, though not written down, is in the spirit of McMullen s earlier result [29] that maximal cusps are dense in the boundary of any Bers slice of quasi-fuchsian space. The main result of this section, Theorem 5.2, asserts that there is a dense G δ -set of purely loxodromic, analytically tame representations in B k. This theorem generalizes and provides an alternate proof of Theorem 8.2 in [14]. The proof of Theorem 5.2 makes use of Proposition 4.5 and McMullen s theorem. We use Proposition 4.5 to show that if ρ B k is well-approximated by an infinite sequence of maximal cusps, then its convex core can be exhausted by nearly isometric copies of the truncated convex cores of the maximal cusps; this implies that ρ is analytically tame. McMullen s theorem guarantees that the G δ -set of points in B k which are well-approximated by an infinite sequence of maximal cusps is dense. We now recall the definition of an analytically tame hyperbolic 3-manifold. Definition: A hyperbolic 3-manifold N with finitely generated fundamental group is analytically tame if C(N) may be exhausted by a sequence of compact submanifolds {M i } with piecewise smooth boundary such that (1) M i M j if i < j, where M j denotes the interior of M j considered as a subset of C(N), (2) M i = C(N), (3) there exists a number K > 0 such that the boundary M i of M i has area at most K for all i, and

22 5. All over the boundary of Schottky space 22 (4) there exists a number L > 0 such that N 1 ( M i ) has volume at most L for every i. While the definition of analytic tameness is geometric in nature, it does have important analytic consequences. In particular, for an analytically tame group Γ one can control the behavior of positive Γ-invariant superharmonic functions on H 3. Specifically, we will make extensive use of the following result, which is contained in Corollary 9.2 of [8]. Proposition 5.1 If N = H 3 /Γ is analytically tame and Λ(Γ) = C then all positive superharmonic functions on N are constant. 5.1 We are now in a position to state Theorem 5.2. Theorem 5.2 For all k 2, there exists a dense G δ -set C k in B k, which consists entirely of analytically tame Kleinian groups whose limit set is the entire sphere at infinity. The proof of Theorem 5.2 involves the following three lemmas. The first lemma is contained in Chuckrow [12]. Lemma 5.3 The set U k of purely loxodromic representations in B k is a dense G δ -set in B k. Moreover, if ρ U k, then Λ(ρ(F k )) = C. 5.3 The second lemma is an adaptation of Bonahon s bounded diameter lemma [5]. Lemma 5.4 For every δ > 0, there is a number c k (δ) with the following property. Let ɛ > 0 be given, let ω be any maximal cusp in D k, and set N = H 3 /ω(f k ). If δ(n) > ɛ, then any two points in D ɛ (N) may be joined by a path β in D ɛ (N) such that β N thick(δ) has length at most c k (δ).

23 5. All over the boundary of Schottky space 23 Proof of 5.4: Let δ > 0 be given. In order to define c k (δ) we consider a hyperbolic 2-manifold P which is homeomorphic to a thrice-punctured sphere; there is only one such hyperbolic 2-manifold up to isometry. Since P thick(δ) is a compact subset of the metric space P, it has a finite diameter d(δ). It is clear that any two points in P may be joined by a path β such that β P thick(δ) has length at most d(δ). We set c k (δ) = (2k 2)d(δ) + 3k 3. Now let ω be any maximal cusp in D k, and set N = H 3 /ω(f k ). We consider an arbitrary component S of C(N). Then S is a totally geodesic thrice-punctured sphere. Hence S, with its intrinsic metric, is isometric to P. Furthermore, the inclusion homomorphism π 1 (S) π 1 (N) is injective, and hence S thin(δ) N thin(δ). It follows that any two points in S may be joined by a path β in S such that β N thick(δ) has length at most d(δ). Now, since D ɛ (N) is a compact core for N and π 1 (N) is a free group of rank k, we see that D ɛ (N) is a handlebody of genus k. In particular, there are exactly 2k 2 components of C(N) and exactly 3k 3 annular components of D ɛ (N) C(N). Also recall, from Lemma 4.1 that each component of D ɛ (N) C(N) has diameter at most 1. Thus, since D ɛ (N) is connected, we see that any two points in D ɛ (N) may be joined by a path β such that β N thick(δ) has length at most (2k 2)d(δ) + 3k 3 = c k (δ). 5.4 In the following lemma, and in the rest of the section, we arbitrarily fix a base point z 0 H 3, and we let X k denote the set of all representations ρ B k such that z 0 lies in the interior of CH(ρ(F k )) relative to H 3. Lemma 5.5 The set X k is an open dense subset of B k. Proof of 5.5: Given a representation ρ X k, we have that z 0 lies in the interior of CH(ρ(F k )). Hence z 0 lies in the interior of some ideal tetrahedron T with vertices in Λ(ρ(F k )). Since the fixed points of elements of ρ(f k ) are dense in Λ(ρ(F k )), we may assume that the vertices of T are attracting fixed points of elements ρ(g 1 ),..., ρ(g 4 ) of ρ(f k ). It follows that for any ρ B k sufficiently close to ρ, the attracting fixed points of ρ (g 1 ),..., ρ (g 4 ) span a tetrahedron having z 0 as an interior point. Hence z 0 lies in the interior of CH(ρ (F k )). This shows that X k is an open subset of B k. If ρ U k, then ρ X k since Λ(ρ(F k )) = C. Hence, Lemma 5.3 implies that X k is dense.

24 5. All over the boundary of Schottky space Proof of 5.2: By the result of McMullen s discussed at the beginning of this section, M k is a dense subset of B k. In view of Lemma 5.5 it follows that M k X k is also dense in B k. For each ω M k X k and each ɛ > 0 we will define a neighborhood V (ɛ, ω) of ω in X k B k. Set N ω = H 3 /ω(f k ). Since ω X k, we have z 0 CH(ω(F k ); in the basepoint convention given in Section 4 we have zω 0 C(N ω). Hence either zω 0 D ɛ (N ω ), or zω 0 lies in the interior of the ɛ-thin part of N ω. Let η(ɛ, ω) = min{ɛ, δ(n ω )}. If zω 0 D ɛ(n ω ) then we set V (ɛ, ω) = U(η(ɛ, ω), z 0, ω) X k, where U(η(ɛ, ω), z 0, ω) is the open set given by Proposition 4.5. If zω 0 lies in the ɛ-thin part of N ω, we take V (ɛ, ω) to be a neighborhood of ω in X k such that for every ρ V the point zρ 0 lies in the interior of the ɛ-thin part of H 3 /ρ(f k ). (Such a neighborhood exists because there is an element g F k such that dist(z 0, ω(g) z 0 ) < ɛ. For any ρ sufficiently close to ω we have dist(z 0, ρ(g) z 0 ) < ɛ.) We now set W k (ɛ) = ω Mk X k V (ɛ, ω). Since M k X k is dense in B k, the set W k (ɛ) is an open dense subset of X k for every ɛ > 0. Since U k is a dense G δ -set in B k it follows that C k = U k W k ( 1 m Z + m ) is a dense G δ -set in X k. In order to complete the proof, we need only to show that each element of C k is analytically tame and has the entire sphere as its limit set. Let ρ : F k PSL 2 (C) be a representation in C k. Set N = H 3 /ρ(f k ). Lemma 5.3 guarantees that C(N) = N. By the definition of C k, for every m Z + there exists a maximal cusp ω m : F k PSL 2 (C) such that ρ V ( 1, ω m m). Let N m = H 3 /ω m (F k ) and let 1 m 0 be a positive integer such that 2m 0 is less than the injectivity radius of N at zρ. 0 In what follows we consider an arbitrary integer m m 0. By the definition of m 0 the point zρ 0 lies in the 1 -thick part of N. By the m definition of the sets V ( 1, ω m m), it follows that ρ U(η( 1, δ(n m m)), z 0, ω m ). Let D m = D η( 1,δ(Nm))(N m). Proposition 4.5 guarantees that there is a map m φ m : D m N, such that

25 5. All over the boundary of Schottky space 25 (1) φ m maps D m homeomorphically onto a manifold with piecewise smooth boundary, and is 2-biLipschitz with respect to the intrinsic metrics on D m and φ m (D m ), (2) φ m (z 0 ω m ) = z 0 ρ, (3) vol N 1 ( (φ m (D m ))) 8β(N m ) = 32πκ(k 1), and (4) φ m ((N m ) thin(δ) D m ) N thin(2δ) for any δ < λ 0 2. (Here κ is the constant given by Lemma 4.3. Notice that since π 1 (N m ) is a free group we have τ(n m ) = 0 and hence β(n m ) = 2πκσ(N m ) = 4πκ(k 1).) We set δ 0 = λ 0 /3 and M m = φ m (D m ). Then by (4) we have M m φ m (D m (N m ) thin(δ0 )) N thin(2δ0 ). Hence, by Lemma 5.4, any two points in M m may be joined by a path β in M m such that β N thin(2δ0 ) has length at most 2c k (δ 0 ). Let r > 0, and let X(r) denote the set of points x N for which there exists a path β beginning in B(r, zρ) 0 and ending at x such that β N thick(2δ0 ) has length at most 2c k (δ 0 ). Since ρ(f k ) is purely loxodromic, each component of N thin(2δ0 ) is compact. Moreover, the components of N thin(2δ0 ) are separated by a distance of at least L(2δ 0 ), so there exist only a finite number of components of N thin(2δ0 ) contained in X(r). Therefore X(r) is compact. Set ζ(r) = min x X(r) inj N (x), and set m 1 = max(m 0, 2/ζ(r)). If m > m 1 then M m B(r, zρ 0) =, since any point in M m may be joined by a path of length at most 2c k (δ 0 ) to a point of injectivity radius 2 m (namely any point in φ m ( D m C(N m ))). Since zρ 0 M m by (2), and since M m B(r, zρ 0) =, we see that B(r, z0 ρ ) M m for every m > m 1. We therefore have m>m1m m = N = C(N). We may pass to a subsequence M mj such that M mj M mj+1 for all j and j Z+ M mj = N = C(N). By (3) we have vol N 1 ( M m ) 32πκ(k 1) for all m > m 1. Moreover, by (1) and Lemma 4.4 we have Thus N is analytically tame. area M m 4α(N m ) = 14πσ(N m ) = 28π(k 1). 5.2

26 6. Free groups and displacements 26 6 Free groups and displacements This section is devoted to the proof of the following theorem, which includes the Main Theorem stated in the introduction. Theorem 6.1 Let k 2 be an integer and let Φ be a Kleinian group which is freely generated by elements ξ 1,..., ξ k. Suppose that either (a) Φ is purely loxodromic and topologically tame, or (b) Φ is geometrically finite, or (c) Φ is analytically tame and Λ(Φ) = C, or (d) the hyperbolic 3-manifold H 3 /Φ admits no non-constant positive superharmonic functions. Let z be any point of H 3 and set d i = dist(z, ξ i z) for i = 1,..., k. Then we have k e d i 2. i=1 In particular there is some i {1,..., k} such that d i log(2k 1). Note that if we had d i < log(2k 1) for i = 1,..., k it would follow that k 1 i=1 1 + e d i > k 1 2k = 1 2. Thus the last sentence of Theorem 6.1 does indeed follow from the preceding sentence. Conditions (a) (d) of Theorem 6.1 are by no means mutually exclusive. In particular, according to Proposition 5.1, condition (c) implies condition (d). The following elementary inequality will be needed for the proof of Theorem 6.1.